rearrange the variable layout of structure tunnel_desc, enable the LTO and make O3 as default level in release build.

4 files changed, 24 insertions, 27 deletions

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 3757893..a5bec6c 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -1,4 +1,4 @@
-cmake_minimum_required(VERSION 3.0)
+cmake_minimum_required(VERSION 3.9)
 
 # Must be set before project, otherwise you cannot choose between gcc and clang
 option(ENABLE_FUZZING_TEST "Enable fuzzing test" False)
@@ -42,10 +42,12 @@ endif()
 set(CMAKE_C_FLAGS "-std=gnu99 -m64 -march=${MACHINE} -fPIC -Wall -Wno-format-truncation")
 set(CMAKE_CXX_FLAGS "-std=gnu++11 -m64 -march=${MACHINE} -fPIC -Wall -Wno-format-truncation")
 
-#set(CMAKE_C_FLAGS_RELEASE "-O3")
-#set(CMAKE_CXX_FLAGS_RELEASE "-O3")
-#set(CMAKE_C_FLAGS_RELWITHDEBINFO "-O3 -g")
-#set(CMAKE_CXX_FLAGS_RELWITHDEBINFO "-O3 -g")
+# enable LTO
+set(CMAKE_INTERPROCEDURAL_OPTIMIZATION True)
+
+# Force -O3 for release and relwithdebinfo
+string(REPLACE "-O2" "-O3" CMAKE_CXX_FLAGS_RELWITHDEBINFO "${CMAKE_CXX_FLAGS_RELWITHDEBINFO}")
+string(REPLACE "-O2" "-O3" CMAKE_C_FLAGS_RELWITHDEBINFO "${CMAKE_CXX_FLAGS_RELWITHDEBINFO}")
 
 if(ENABLE_WARNING_AS_ERROR)
 	set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -Werror")
diff --git a/infra/src/vnode_common.c b/infra/src/vnode_common.c
index eed2cd9..c49faf0 100644
--- a/infra/src/vnode_common.c
+++ b/infra/src/vnode_common.c
@@ -55,7 +55,6 @@ static struct tunnel_desc * tunnel_new(const char * symbol, unsigned int sz_excl
         goto errout;
     }
 
-    snprintf(desc->symbol, sizeof(desc->symbol), "%s", symbol);
     desc->tunnel_size = sz_exclusive + sz_shared;
 
     desc->en_buffer = ZMALLOC(sizeof(void *) * MR_LIBVNODE_MAX_SZ_BURST);
@@ -97,13 +96,13 @@ errout:
 /* Delete a tunnel */
 static int tunnel_delete(struct tunnel_desc * desc)
 {
-	for (int i = 0; i < desc->sz_en_buffer_used; i++)
+    for (int i = 0; i < desc->sz_en_buffer_used; i++)
     {
         rte_pktmbuf_free(desc->en_buffer[i]);
     }
 
-	struct rte_mbuf * mbuf;
-	while (rte_ring_dequeue(desc->tunnel_object, (void **)&mbuf) == 0)
+    struct rte_mbuf * mbuf;
+    while (rte_ring_dequeue(desc->tunnel_object, (void **)&mbuf) == 0)
     {
         rte_pktmbuf_free(mbuf);
     }
@@ -115,11 +114,12 @@ static int tunnel_delete(struct tunnel_desc * desc)
         rte_pktmbuf_free(desc->rt_buffer[i]);
     }
 
-    MR_VERIFY_2(rte_ring_empty(desc->tunnel_object) == 1, "Tunnel %s is not empty", desc->symbol);
+    MR_VERIFY_2(rte_ring_empty(desc->tunnel_object) == 1, "Tunnel %p is not empty", desc);
     rte_free(desc->en_buffer);
     rte_free(desc->rt_buffer);
     rte_ring_free(desc->tunnel_object);
     rte_free(desc);
+
     return 0;
 }
 
diff --git a/infra/src/vnode_common.h b/infra/src/vnode_common.h
index 4dc35b9..061bc3b 100644
--- a/infra/src/vnode_common.h
+++ b/infra/src/vnode_common.h
@@ -23,36 +23,24 @@ struct tunnel_desc
 {
     /* first cacheline, read only */
     RTE_MARKER cacheline0;
-
-    /* Tunnel Name */
-    char symbol[MR_SYMBOL_MAX];
     /* Tunel Object, real object to hold pointers */
     struct rte_ring * tunnel_object;
     /* Tunnel Size */
     unsigned int tunnel_size;
 
-    /* Tunnel Enqueue Buffer */
+    unsigned int sz_en_buffer;
     struct rte_mbuf ** en_buffer;
+
+    unsigned int sz_rt_buffer;
     struct rte_mbuf ** rt_buffer;
 
-    /* second cacheline, read/write */
+    /* second cache line, read/write on prod cores */
     RTE_MARKER cacheline1 __rte_cache_min_aligned;
-    int32_t inflight_credits;
 
-    /* Tunnel Enqueue Buffer Size */
-    unsigned int sz_en_buffer;
     /* Tunnel Enqueue Buffer Used */
     unsigned int sz_en_buffer_used;
-    /* Return Buffer Size */
-    unsigned int sz_rt_buffer;
     /* Return Buffer Used */
     unsigned int sz_rt_buffer_used;
-
-#if VNODE_CHECK_THREAD_SAFE
-    /* For debug, to check concurrent access of en_buffer */
-    rte_spinlock_t lock_thread_safe_check;
-#endif
-
     /* counters */
     uint64_t on_line;
     uint64_t deliver;
@@ -62,6 +50,14 @@ struct tunnel_desc
     /* queue length */
     unsigned int q_len;
     float q_len_avg;
+
+    RTE_MARKER cacheline2 __rte_cache_min_aligned;
+    int32_t inflight_credits;
+
+#if VNODE_CHECK_THREAD_SAFE
+    /* For debug, to check concurrent access of en_buffer */
+    rte_spinlock_t lock_thread_safe_check;
+#endif
 };
 
 struct tunnel_block
diff --git a/service/CMakeLists.txt b/service/CMakeLists.txt
index cac5fc2..ea427fe 100644
--- a/service/CMakeLists.txt
+++ b/service/CMakeLists.txt
@@ -13,7 +13,6 @@ target_link_libraries(mrzcpd MESA_prof_load_static libdpdk ${SYSTEMD_LIBRARIES})
 target_link_libraries(mrzcpd rt pthread dl infra z elf)
 target_include_directories(mrzcpd INTERFACE "${CMAKE_CURRENT_SOURCE_DIR}/include/")
 target_include_directories(mrzcpd INTERFACE ${SYSTEMD_INCLUDE_DIRS})
-
 install(TARGETS mrzcpd RUNTIME DESTINATION ${MR_INSTALL_BINDIR} COMPONENT Program)
 
 add_executable(TestOLP test/TestOLP.cc src/olp_6500.c)